f265c87e01
When a foreign event loop that does not enter an alertable wait state is running (which is also the case when a native dialog window is modal), pipe handlers would freeze temporarily due to their APC callbacks not being invoked. We address this problem by moving the I/O callbacks to the Windows thread pool, and only posting completion events to the main loop from there. That makes the actual I/O completely independent from any main loop, while the signal delivery works also with foreign loops (because Qt event delivery uses Windows messages, which foreign loops typically handle correctly). As a nice side effect, performance (and in particular scalability) is improved. Several other approaches have been tried: 1) Using QWinEventNotifier was about a quarter slower and scaled much worse. Additionally, it also required a rather egregious hack to handle the (pathological) case of a single thread talking to both ends of a QLocalSocket synchronously. 2) Queuing APCs from the thread pool to the main thread and also posting wake-up events to its event loop, and handling I/O on the main thread; this performed roughly like this solution, but scaled half as well, and the separate wake-up path was still deemed hacky. 3) Only posting wake-up events to the main thread from the thread pool, and still handling I/O on the main thread; this still performed comparably to 2), and the pathological case was not handled at all. 4) Using this approach for reads and that of 3) for writes was slightly faster with big amounts of data, but scaled slightly worse, and the diverging implementations were deemed not desirable. Fixes: QTBUG-64443 Change-Id: I66443c3021d6ba98639a214c3e768be97d2cf14b Reviewed-by: Oswald Buddenhagen <oswald.buddenhagen@gmx.de> |
||
|---|---|---|
| .. | ||
| auto | ||
| baselineserver | ||
| benchmarks | ||
| global | ||
| libfuzzer | ||
| manual | ||
| shared | ||
| testserver | ||
| CMakeLists.txt | ||
| README | ||
This directory contains autotests and benchmarks based on Qt Test. In order
to run the autotests reliably, you need to configure a desktop to match the
test environment that these tests are written for.
Linux X11:
* The user must be logged in to an active desktop; you can't run the
autotests without a valid DISPLAY that allows X11 connections.
* The tests are run against a KDE3 or KDE4 desktop.
* Window manager uses "click to focus", and not "focus follows mouse". Many
tests move the mouse cursor around and expect this to not affect focus
and activation.
* Disable "click to activate", i.e., when a window is opened, the window
manager should automatically activate it (give it input focus) and not
wait for the user to click the window.